A dynamic generation approach for ensemble of extreme learning machines

Abstract

Extreme learning machine (ELM) as one new learning algorithm has been proposed for single hidden-layer feed-forward neural network (SLFN). In contrast with the popular back-propagation (BP) algorithm, ELM often has obviously faster learning speed and stronger generalization performance. However, ELM lacks stability as the weights and biases between the input layer and the hidden layer are randomly assigned, and meanwhile, it often suffers from overfitting as the learning model will approximate all training instances well. In this article, a dynamic generation approach for ensemble of extreme learning machine (DELM) is proposed to overcome the problems above. Specifically, cross-validation and one target function are embedded into the learning phase. Experimental results on several benchmark datasets indicate that DELM is robust and accurate.